the impact of mobile money on poor rural households · 2019-06-25 · clusters had airtel money...
TRANSCRIPT
Policy Research Working Paper 8913
The Impact of Mobile Money on Poor Rural Households
Experimental Evidence from Uganda
Christina WieserMiriam Bruhn
Johannes KinzingerChristian Ruckteschler
Soren Heitmann
IFC-Mastercard Foundation Partnership for Financial InclusionJune 2019
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Produced by the Research Support Team
Abstract
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Policy Research Working Paper 8913
This paper studies the effect of rolling out mobile money agents in rural Northern Uganda. In a randomized experi-ment, 168 areas were randomly selected to receive an agent in 2017, with another 163 areas serving as a control group. Administrative data on mobile money transactions suggest that the agent rollout increased the probability of send-ing and receiving peer-to-peer transfers. Data from a 2018 survey of more than 4,500 households show that the agent
rollout led to cost-savings for remittance transactions. It also doubled the nonfarm self-employment rate, from 3.4 to 6.4 percent, and reduced the fraction of households with very low food security from 62.9 to 47.2 percent, in areas far from a bank branch. The analysis finds no effect on savings, agricultural outcomes, or poverty. Overall, the findings add new evidence that mobile money can improve livelihoods even in poor and remote settings.
This paper is a product of the IFC-Mastercard Foundation Partnership for Financial Inclusion. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at [email protected] and [email protected].
The Impact of Mobile Money on Poor Rural Households: Experimental Evidence
from Uganda1
Christina Wieser, Miriam Bruhn, Johannes Kinzinger, Christian Ruckteschler, Soren Heitmann
JEL Classification: C93; E42; G23; O12; O18.
Keywords: Mobile money, remittances, food security.
1 This study was funded by the Bill and Melinda Gates Foundation. It is implemented through and draws important experiences from the IFC‐Mastercard Foundation Partnership for Financial Inclusion (“the Partnership”), which leverages IFC’s expanding Mobile Network Operators client networks in Africa and its emerging expertise in mobile financial services to catalyze innovative and low‐cost approaches to expand financial services to low‐income populations. We thank Siobhan Murray from the World Bank for her help with mapping and defining the clusters of enumeration areas used in this study; Richard Chamboko from the IFC for his comments and edits; and Richard Yego, Acting Director, Airtel Money, Uganda, for his operational support to control the agent roll‐out and implement the study.
2
1. Introduction
Digital Financial Services (DFS) have spread rapidly in developing countries, enabling mobile phone users
to make financial transactions on their phone, such as transferring money, buying airtime, and paying
bills. Recent evidence suggests that mobile money has allowed households to smooth consumption in the
face of shocks (Jack and Suri 2014) and that it has increased per capita consumption levels and decreased
poverty in the long‐run (Suri and Jack 2016; Munyegera and Matsumoto 2016).
The most commonly documented channel through which mobile money can affect consumption is
increased remittance transactions and risk sharing. Jack and Suri (2014) show that M‐PESA users in Kenya
are more likely to receive remittances (particularly after experiencing shocks) and these come from more
different senders (see also Jack, Ray, and Suri 2013). Blumenstock, Eagle and Fafchamps (2016) find that
Rwandans used the mobile phone network to transfer airtime to those affected by an earthquake.2
Munyegera and Matsumoto (2016) show that mobile money adopters in rural Uganda receive more and
larger remittances than non‐users.
We study the effect of mobile money in a setting where remittances are less prevalent than in previous
studies: households in rural areas of Northern Uganda, where only 15 percent of households report
receiving remittances compared to 40 percent in Jack and Suri (2014) and 65 percent in Munyegera and
Matsumoto (2016). Apart from increasing remittance transactions, mobile money can potentially benefit
households through other channels. For example, Sekabira and Qaim (2017) find that mobile money can
increase the income of coffee farmers in Central Uganda by facilitating transactions with new buyers. The
areas we study tend to be poorer than those in Sekabira and Qaim (2017)3 and most agricultural
households tend to focus on subsistence farming.
Apart from high poverty rates, areas in our sample also have low access to financial services, with the
median distance to a bank branch being 25.2km. In this context, we ask whether and how rolling out
mobile money agents affects households’ financial behavior, occupational choices, food security, and
poverty rates. Several studies have found that bringing financial services, bank accounts in particular to
rural areas in developing countries can increase savings, business investment and income.4 In contrast,
Dupas et al. (2018) found households in their study in Malawi and Central Uganda were simply too poor
to save in bank accounts. In their setting, the high transaction costs of having to go to a bank branch may
also have played a role; so branchless banking such as through mobile money linked accounts may have
higher impacts on savings (Suri and Jack 2016).
The underlying question in this study is whether the areas in our sample are too remote and poor to
benefit from mobile money (since they receive few remittances and may not have enough income to save)
2 Rwanda did not have a functioning mobile money system at the time, hence the focus on transferring airtime. These findings are also in line with Yang and Choi (2007) who document that Philippine households receive more overseas remittances when they experience income shocks, so that those remittances could be used to smooth consumption. 3 The proportion of the population living in poverty in Uganda increased from 19.7 percent in 2012/13 to 21.4 percent in 2016/17, an equivalent of about 10 million people living below the poverty line. The Northern region is among the poorest regions in the country with poverty rates decreasing from 43.7 to 32.5 percent between 2012/13 and 2016/17 (UBOS, 2018). 4 See Burgess and Pande (2005) and Young (2015) for India, Bruhn and Love (2014) for Mexico, Dupas and Robinson (2013) for Kenya, Brune et al. (2016) for Malawi, and Prina (2015) for Nepal.
3
or whether mobile money can help households to increase remittance receipts and/or increase savings in
mobile accounts. If so, whether increased financial flows into rural areas have measured impact on
beneficiary livelihoods. Another potential benefit of mobile money in our remote sample is that it implies
non‐negligible cost savings for financial transactions. For example, households who receive remittances
typically paid 4,000 Ugandan shillings (US$1) per transaction in transportation costs, which in our study
area represents about 10 percent of per capita daily household expenditures.
To measure the effect of mobile money in poor and remote areas, the International Finance Corporation
(IFC) collaborated with Airtel Uganda to implement a field experiment in Northern Uganda covering a
geographic area of approximately 125,000km2 where 334 clusters of enumeration areas (EAs) were
randomly assigned to a treatment or a control group, stratified by distance to a bank branch. None of the
clusters had Airtel Money agents at baseline. In the treatment group, Airtel Money agents were rolled out
in 2017, with 46 percent of clusters receiving at least one agent.
Data on about 4,000 households from a 2018 follow‐up survey show that the agent rollout increased the
percentage of households using mobile money by about 4.2 percentage points. That is, the agent rollout
increased the percentage of households using mobile money from 12.6 percent in the control group to
16.8 percent in the treatment group, in clusters farther away than 25.2km from a bank branch,5 with no
effect in areas closer to a bank branch.
However, in areas far from a bank branch, we find that those who receive or send remittances were less
likely to report high costs in treatment clusters than in control clusters, after the rollout. These decreased
costs reflect lower transportation costs, as those who received remittances in treatment clusters were 13
percentage points less likely to pay transportation costs, such as motorcycle or mini‐bus taxi fees,
associated with the transaction. That is, in the follow‐up survey, 82 percent of remittance receivers in
treatment areas far from a bank branch reported not paying transportation costs for the transaction,
compared to 69 percent not paying transportation costs in the control group. In line with the previous
literature, administrative data from Airtel on mobile money transactions also suggest that the agent
rollout increased the probability of sending and receiving peer‐to‐peer (P2P) transfers.
In contrast to Suri and Jack (2016) we find no effect of the agent rollout on savings. We also do not see an
effect on the probability of owning livestock or agricultural outcomes, including the amount spent on
seeds or fertilizer use. However, in areas far from a bank branch, the agent rollout statistically significantly
increased the fraction of survey respondents who work in non‐farm self‐employment, by 3 percentage
points. That is, in the follow‐up survey, 6.4 percent of households in treatment areas far from a bank
branch reported working in non‐farm self‐employment, compared to 3.4 percent in the control group. In
line with this finding, we observe that the agent rollout changed the way in which households respond to
negative shocks, such as a drought or flood. The rollout increased the probability of taking work to cope
with the shock and it reduced the probability of changing the household’s diet. Further, consistent with
this result, we find a strong effect of the agent rollout on food security. In areas far from a bank branch,
the agent rollout lowered the probability of having to reduce the number of meals in the last 7 days by
11.8 percentage points. That is, in treatment areas far from a bank branch, 36.8 percent of households
reported having to reduce meal, compared to 48.6 percent in the control group. However, we do not find
5 We picked 25.2km to look at heterogenous treatment effects in the analysis since it is the median distance to a bank branch across the clusters in our sample.
4
an effect of the agent rollout on our measure of poverty, which is computed based on the SWIFT (Survey
on Well‐being via Instant and Frequent Tracking) imputation approach.
Our finding that the agent rollout increased non‐farm self‐employment mirrors the results in Sekabira and
Qaim (2017) who also conclude that mobile money increased non‐farm income. Sekabira and Qaim do
not pin down the channel through which mobile money increased non‐farm income. The channel in our
setting is likely that households used their increased P2P transfer receipts and cost savings from
remittance transactions to invest in self‐employment, which in turn generated income that raised food
security. The fact that most of our statistically significant effects come from clusters far from a bank
branch supports the argument that transaction cost savings contribute to the effect.
Thus, overall, we find that mobile money stimulates self‐employment and increases food security in rural
Northern Uganda. We conclude that mobile money services can improve livelihoods even in very poor
and remote areas.
The rest of this paper proceeds as follows. Section 2 provides background information on mobile money
in Uganda. Section 3 discusses the study design and agent rollout. Section 4 describes the data and
methods. Section 5 presents the results and Section 6 concludes.
2. Mobile Money in Uganda
According to a recent study, 70.9 percent of Ugandans owned a mobile phone; with higher penetration
rates in urban (78.5 percent) compared to rural areas (65.7 percent) in 2017 (CIPESA 2018). Most mobile
phone owners (73.7 percent) used MTN as their service provider, followed by Airtel (61.4 percent) and
Africell (7.3 percent) (multiple selections possible).
Our study focuses on rural Northern Uganda, where penetration of mobile phones is much lower. Data
from our follow‐up survey, conducted in early 2018 show only 27.6 percent of households own a mobile
phone. However, we observe that people tend to share or use someone else’s phone since a higher
proportion (39.9 percent) of households reported that they have access to a mobile phone. Our study also
shows that the most popular mobile phone provider in rural Northern Uganda is MTN. Conditional on
access to a mobile phone, 74.7 percent of households use MTN, 7.9 percent use Airtel, and 9.9 percent
use Africell as their service provider. Airtel’s presence in rural areas of Northern Uganda thus falls short
compared to their national presence (table 1).
In 2009, MTN launched its mobile money platform and it has since then dominated the mobile money
market in Uganda. However, several competitors have entered the market, including three mobile
network operators (MNO) – Airtel, Africell, and Uganda Telecom through M‐Sente – and a few non‐MNO
mobile payment providers, such as M‐Cash, Ezee Money, and Micro‐pay.
Mobile financial services have seen sharp growth in recent years, reaching 22.9 million registered
customers in June 2017 being served by 147,146 agents and accumulating into 1,111 million transactions
and 52.8 trillion Ugandan shillings (approx. 14.6 billion USD) in transaction value from June 2016 to June
2017. Nationwide, 62.1 percent of mobile phone owners sent or transferred money using electronic
methods (Bank of Uganda 2017). Northern Uganda lags behind the rest of the country in terms of
5
availability and use of digital financial services with less than 3 percent of all agents operating in Northern
Uganda, the majority of which (76 percent) were MTN mobile money agents.6
For this study, IFC collaborated with Airtel to roll‐out Airtel Money agents in Northern Uganda in
geographical areas not yet served by Airtel Money. Airtel Money was launched in January 2012, and its
agents offer a variety of services including cash withdrawals, cash deposits, purchase of airtime, sending
and receiving money, bill payment (such as school fees and utilities), and payment for goods and services.
Despite Airtel Money’s ambition to increase the delivery of financial services to the Ugandan population,
growth in Northern Uganda has been slow. IFC has had an interest in increasing Airtel’s mobile money
market share by improving and strengthening its operations for two reasons: (i) to increase the outreach
of mobile money into remote communities and (ii) to strengthen the competitive environment of mobile
money in Uganda, avoiding a monopolistic environment, with active competition, and to improve pricing,
sharing of agents and the potential for interoperability of mobile money providers.
3. Study Design and Agent Rollout
3.1 Study Design
The study was conducted in rural areas in the North of Uganda, covering the three regions West Nile, Mid‐
North and Karamoja. The Ugandan Bureau of Statistics (UBOS) helped to define the study sample. As a
starting point, all rural enumeration areas in all sub‐counties in the West Nile, Mid‐North and Karamoja
regions were used. From this list, sub‐counties that had an Airtel Money agent present were dropped,
based on an agent database provided by the Bill and Melinda Gates Foundation. A random sample of
1,200 enumeration areas (EAs) was drawn from all rural enumeration areas in the remaining 249 sub‐
counties using population proportional to size sampling. The sample was implicitly stratified by sub‐county
to ensure that it included enumeration areas from all sub‐counties.
In September and October 2015, a household listing exercise was conducted in the EAs selected for the
study to generate a sampling frame that could be used for the baseline survey. At the same time, all
businesses in each EA were listed to gain information on potential mobile money agents in the EA. To
minimize potential spillovers of agents to control group EAs, the selected EAs were mapped and grouped
into clusters based on proximity. A 0.5km buffer was drawn around the boundary of each EA and EAs
whose buffers overlapped were grouped. Clusters were thus at least 1km apart from each other.
In this process, some EAs were dropped from the sample since either the listing exercises could not be
conducted (for logistical or security reasons), the listing did not yield any businesses and thus no potential
agents, or the maps were missing from the software to group EAs into clusters. As a result, there were
929 EAs that were grouped into 658 clusters.
Using a computer assisted stratified randomization approach, 329 clusters were assigned to a treatment
group and 329 clusters to a control group. The treatment clusters formed the list of EA clusters where
Airtel Money agents were to be rolled out. No such agents were to be rolled out in control clusters.
The randomization strata were based on three variables. First, a variable equal to one if the cluster
included more than one EA (which was the case for 16 percent of clusters) and equal to zero if the cluster
6 Based on an agent map provided by the Bill and Melinda Gates foundation, available at: http://www.fspmaps.com/#/Uganda/finance.
6
included only one EA. Second, a variable equal to one if the distance to the nearest bank branch was
greater than the median distance across all clusters (25.2km) and equal to zero otherwise. These bank
branches refer to Airtel partner banks and the distance was calculated as the closest distance from any
point on the cluster boundary. The main reason for stratifying on this variable was that Airtel raised
concerns as to whether agents that are further away from these banks can operate well since they may
have a harder time maintaining liquidity. On the other hand, distance to an Airtel partner bank is likely
also correlated with distance to urban centers and other financial services, so that clusters that are further
away from a branch may see greater benefits from obtaining an agent. The third stratification variable
was based on a strategic priority rating provided by Airtel, ranging from 1 to 4. In clusters with greater
priority, more time was to be spent on rolling out agents than in areas with low priority. In total, 16 strata
were created.
3.2 Agent Rollout
The Airtel Money agent rollout took place in treatment clusters between January and June 2017. A
professional services firm was hired to assist with identifying potential agents.7 This firm also helped
agents with the logistics of signing‐up to become Airtel Money agents and provided them with the
necessary equipment, training, and marketing materials. The research team closely supervised the
professional services firm and quality‐checked information provided by them.
During the roll‐out, the professional services firm activated 400 agents who undertook a successful first
transaction. To further assess the extent of activity among these agents, we matched the phone numbers
associated with these agents with Airtel Money transactions data. The records provided by the
professional services firm do not fully line up with the Airtel data: only 370 of the 400 agents were found
in the Airtel data, which could potentially be due to typos in the recorded phone number.
Based on the matched 370 phone numbers, panel A of figure 1 shows the number of active agents in the
Airtel Money transactions data per month from December 2016 to November 2017. The number of active
agents steadily increased, reaching a high of 285 in August 2017 with a relatively stable number of agents
through November 2017.8 Usage for the active agents grew steadily as can be seen in panel B of Figure 1.
The total number of transactions across all agents reached over 24,000 transactions, with an average of
70 transactions per agent per month (conditional on having any transactions that month).9
Although the agent rollout specifically targeted treatment areas, the implementation was not exact. The
administrative records from the agent roll‐out list the EA cluster where the agent is located. To verify
these locations, we used GPS coordinates that were available for 375 agents. About 45 percent of GPS
locations are inside a treatment cluster and another 39 percent were within 1km of a treatment area, with
7 As a starting point to identify agents, information from the 2015 business listing exercise was used. The following information was collated into an index: (i) Number of employees; (ii) Type of business (service, retail, others); (iii) Annual turnover; (iv) Proximity to closest bank branch; (v) Business closed when visiting bank branch; and (vi) Business used mobile money before. In treatment areas with more than one listed business, the business with the highest index score was selected to be the first business to be approached. 8 Not all 370 rolled‐out agents transacted at the same time in any of the months, therefore the highest number of agents shown for August 2017 is lower than 370. 9 Agent transactions in the transaction database include cash‐in and cash‐out transactions, peer‐to‐peer transfers, and top ups.
7
a total of 94 percent falling within 2km of a treatment area. Two GPS locations were inside a control area,
about 2 percent were within 1km of a control area and 13.6 percent fell within 2km of a control area.
Agents were not evenly distributed across clusters. Among the 161 treatment group clusters with a low
priority rating from Airtel (a rating of 1 or 2), only 24 percent of clusters received at least one agent during
the rollout. None of the 163 control group clusters with a low priority rating received an agent. Among
the 168 treatment clusters with a high priority rating (3 or 4), 46 percent of clusters received at least one
agent. Close to 2 percent of the 166 control clusters with a high priority rating also received an agent.
Since the rollout resulted in such low agent penetration for low priority clusters, the impact analysis
focuses only on high priority clusters. Given this, the 168 low priority clusters were dropped from the
sample and only the 334 high priority clusters were retained. Our results here thus apply to areas that
were deemed high priority based on Airtel’s business strategy. Figure 2 shows the 334 high priority
clusters.
According to 2015 data provided by the Bill and Melinda Gates Foundation, in the study sample of 334
clusters, only three clusters had a mobile money agent before the agent rollout. All three were MTN
agents. For 39 clusters, an agent was located within 1km of the cluster border (again mostly MTN agents),
but for the remaining 295 clusters, the closest agent was more than 1km away, with an average distance
of 8km to the closest agent. Figure 3 shows the Gulu district as an example, illustrating the location of
treatment and control clusters, along with pre‐existing agents and Airtel agents that were rolled out as
part of this study.
4. Data and Methods
4.1 Survey Data
We use panel data collected in two survey rounds from randomly selected households in rural Northern
Uganda. Based on the household listing exercise conducted in September and October 2015 (see section
3.1), we randomly selected 8 households in each EA for a baseline survey to be conducted in December
2015 and January 2016, giving a total sample of 3,720 households across the 334 clusters (466 EAs)
included in the analysis.10
In the follow‐up survey conducted two years after the baseline, in January and February 2018, we
interviewed 9 households in each EA, increasing the sample size to a total of 4,541 households.11 We
10 For each EA, 8 “original households” and 7 “substitute households” were randomly drawn from the household list collected during the listing exercise. During fieldwork, interviewers were provided with a list of only the 8 original households to be interviewed in each EA. The survey interviewed household heads or household members aware or involved in making financial decisions within the household. In cases where such a respondent was not available, the interviewer made 3 call‐backs to the household before it was replaced with a substitute household from the substitution list. In cases were the household declined to participate or had moved to another location, the interviewer substituted that household without making 3 call‐backs. Overall, 91.3 percent of the households interviewed were original households while 8.7 percent were substitute households. 11 The baseline survey had also interviewed households in EAs that were dropped prior to randomization for the reasons described in section 3. Instead of re‐interviewing households in the dropped EAs, we increased the sample size in the 929 EAs included in the randomization. For each EA, field supervisors were provided with a list of 8 households interviewed during the baseline as well as one or two additional households randomly drawn from the household listing. In addition, they were provided with 15 randomly drawn substitution households. The larger
8
aimed to re‐interview baseline households in the second round. Of the 3,720 households interviewed in
the baseline survey, we re‐interviewed 3,082 households during the follow‐up survey, implying an
attrition rate of 17 percent (equally distributed in the treatment and control groups).12 Households that
could not be re‐interviewed were replaced with other randomly sampled households in the same EA.
Additional households (1 per EA) were also sampled randomly in the same locations.
In both survey rounds, we conducted face‐to‐face interviews. The questionnaire focused on details of
sociodemographic details of households, labor outcomes of the respondents, information on usage of
mobile phones and mobile money, financial transactions, as well as food security and shocks. Some of
these questions were added in the follow‐up survey and were thus not available at baseline.
4.2 Airtel Transactions Data
We obtained monthly data from Airtel on mobile money transactions for the years 2016 and 2017 (going
up to November 2017), including seven types of transactions: sending a P2P transfer, receiving a P2P
transfer, cash‐in, cash‐out, bill pay, airtime top‐up and data top‐up.
To analyze these data, we had to map them to our study clusters, which was not a straightforward
exercise. We have the phone number associated with each transaction, however, we do not have
information on who owns the phone number or where the owner is located. The location can be
approximated by the location of the cell phone tower the number uses the most. However, the radius of
cell phone towers is rather large (up to 40km) which means that it could cover multiple clusters in our
study.
We thus mapped the transaction data to our study locations by using the phone numbers households
reported in our baseline and follow‐up surveys. That is, we complied a list of all phone numbers ever
reported by the households in our follow‐up survey and merged these data with the call records. We then
sum the number of transactions in each category across all phone numbers associated with a study cluster
to get the total number of transactions per cluster. Taking all categories together, this gives us only 2.4
transactions per cluster per month, on average. Since the agent data discussed in section 3.2 suggest that
not all agents transact each month, we further sum the data across months, to obtain yearly totals for
2016 and 2017.
Table 2a displays the average number of transactions in the different categories for our 334 study clusters
in 2016 and 2017, with all transactions combined, showing an average of 15.5 in 2016 and 24.9 in 2017.
The table also shows that less than 15 percent of clusters had any transactions each year. In the impact
analysis below, we thus use dummy variables as outcome variables, where the dummy variables are equal
to one if a cluster had any transactions, in a given category, and equal to zero otherwise. Table 2b shows
that these dummy variables were not statistically significantly different across treatment and control
clusters before the agent rollout took place, i.e. in 2016.
number of substitution households provided reflects the expectation that more households might have moved or dissolved due to the two‐year time gap since the household listing. 12 Around 90 percent of all attrition cases were due to the household having left the enumeration area since the baseline survey was conducted. The remaining cases were the result of respondents being temporarily unavailable or refusing to participate, or the household having dissolved (e.g. because of the death of the household head).
9
4.3 SWIFT Poverty Measure
To gauge and understand households’ welfare trends and the impact of mobile financial services on poor
rural households, the SWIFT approach was used. SWIFT is a rapid poverty assessment tool. It produces
accurate household expenditure and poverty data in a timely, cost‐effective and user‐friendly manner. It
has been used to improve availability of official poverty statistics and monitoring welfare impact of
projects prepared by various Global Practice Groups of the World Bank and IFC. Compared with typical
household consumption data collection in which household consumption or income data are collected via
a complex and long consumption module, SWIFT collects information on poverty correlates, which are
then used to impute household expenditure via a survey‐to‐survey imputation technique. Subsequently,
poverty estimates from the imputed expenditure data are calculated.13
4.4 Impact Estimation
We estimate the impact of the Airtel Money agents with the following intention‐to‐treat (ITT) equation
y , , α βAgent , ∑ γ d δMale , , ε , , (1)
where y , , , is a follow‐up survey measure of mobile money usage, financial behavior, or other outcome
of household i, in cluster c and randomization strata s. The variable Agent , indicates whether the cluster was randomly selected for the agent rollout and is thus equal to one for the treatment group and equal
to zero for the control group. We control for randomization strata dummies d , as well as for whether the
household head is male Male , , since table 2a shows that the fraction of male household heads is slightly
larger in the treatment group than in the control group. When data on the outcome were collected at
baseline, we also control for the baseline value of the outcome variable. For observations where the
baseline value of the outcome is missing (because the household was not interviewed at baseline or due
to nonresponse), we replace this value with zero and include a dummy variable indicating that the value
is missing, to keep the observation in the sample. Standard errors are clustered at the EA cluster level.
The main coefficient of interest is β, which represents the treatment effect of the cluster being selected
for the agent rollout. The coefficient β in equation 1 is also equal to the difference in means of the
outcome variable y , , , across the treatment and control groups, conditional on strata dummies and
household head gender.
In heterogenous treatment effect regressions, we examine whether the effect of the agent rollout varies
by distance to a bank branch. As described in section 3, we stratified the randomization by whether the
distance from the cluster to the nearest bank branch was greater than 25.2km or not. In equation 1 we
replace the treatment variable Agent , with two interaction terms, as follows
13 The basic idea of the SWIFT methodology is to first create consumption models based on a national household survey in which both consumption data and data on poverty correlates are available and reliable. The model identifies the most closely correlated variables to poverty estimates. We then identify the corresponding survey questions to the variables included in the model and collect data on these variables. The marginal cost of adding these, typically 10‐15, survey questions is negligible. Once the consumption model is obtained, the Multiple Imputation method that performs repeated drawings on different random components as well as coefficients was employed to detect the accuracy of prediction and stability of the model over time.
10
y , , α β Agent , ∗ Far , β Agent , ∗ Near , ∑ γ d δMale , , ε , , (2)
where Far , is a dummy variable equal to one if the distance from the cluster to the nearest bank branch
was greater than 25.2km and equal to zero otherwise. Similarly, Near , is a dummy variable equal to one
if the distance from the cluster to the nearest bank branch was smaller than or equal to 25.2km and equal
to zero otherwise. The coefficient β represents the effect of the agent rollout for clusters that are more
than 25.2km away from a bank branch, while β gives the effect of the agent rollout for clusters that
are less than or equal to 25.2km away from a bank branch.
5. Results
5.1 Descriptive Statistics
Table 3a shows summary statistics. Column 2 reports the sample size, which can vary from question to
question due to nonresponse and since some questions were only asked conditional on answers to other
questions. Columns 2 through 4 report averages and standard deviations separately for the control group
and then the treatment group. Column 5 shows the p‐value of the difference in treatment and control
group means, conditional on strata dummies.
Since the sample at follow‐up included some households that were different from the baseline sample,
panel A in table 3a shows data from the baseline survey, and then panel B shows data from the follow‐up
survey to check whether background characteristics were balanced across the treatment and control
groups in the follow‐up sample. The background data from the follow‐up survey look very similar to those
from the baseline survey.
Overall, the control and treatment group means are close to each other for all variables, as can be
expected due to random assignment. The p‐values are higher than 0.1 for all except one variable, implying
that the differences between treatment and control group means are, for the most part not statistically
significant at conventional levels. The only variable that shows a statistically significant difference is the
dummy variable for whether the household head is male or not. In the treatment group, 75 percent of
household heads are male, as opposed to 71 percent in the control group. To ensure that this difference
does not influence our results, we control for gender of the household head in our impact analysis, as
described in section 4.
The background characteristics listed in panel A of table 3a show that household heads were on average
42 years old at baseline. Only about 45 percent of households had at least one member with completed
primary education. The average household size was 5.4. In terms of phone and mobile money usage, only
about 45 percent of respondents had access to a phone and 14 percent report that they used mobile
money in the last 3 months. Although close to 80 percent of respondents reported saving some money,
only 6 percent saved in a mobile money account during the last 6 months.
The surveys included modules on remittances, asking if households had received money from or sent
money to somebody outside the household during the past 6 months, since the previous literature has
documented that mobile money can facilitate remittance transactions (see Jack, Ray, and Suri 2013). Only
15 percent of respondents reported that they received money from somebody outside the household,
11
but in about 45 percent of cases they received the transfer via mobile money (this percentage increased
to almost 75 percent in the follow‐up survey).14
For households who received money, the surveys asked for details of the transaction only for the person
they received money from most frequently. About half of the households who received money reported
receiving it more than once and about one‐third of these households reported receiving more than
100,000 shillings (approximately US$25). Similarly, about half of the households who sent money reported
sending it more than once and one‐third reported sending more than 100,000 shillings. About 30 percent
of households who received money and 25 percent of households who sent money report paying no fees.
Average fees for receiving or sending money were around 3,500 shillings, but this variable has a large
standard deviation and includes some outliers. The median fee was 2,000 shillings. About half of the
households receiving or sending money paid no transportation costs (because they walked or biked). The
average amount paid by those who paid transportation costs was around 5,500 shillings. The higher means
in some groups reflect outliers. The median transportation cost was 4,000 shillings.
Households that had either received or sent money were asked which problems they encountered when
making the transaction. Table 3a shows some frequently mentioned problems: about 30 percent of
respondents reported high prices, 20 percent said the agent was far, 20 percent encountered a technical
problem, and about 10 percent reporting being charged the wrong amount.
Close to two‐thirds of respondents reported that they have a job that earns income. Most of these jobs
were in farming. Only 11 percent of the sample worked outside farming, split almost evenly between wage
employment and self‐employment. That is, about 5 percent were self‐employed doing work other than
farming.
Finally, since the previous literature has found that mobile money can help households cope with shocks
and smooth consumption (Blumenstock, Eagle and Fafchamps 2016; Jack and Suri 2014), the survey
included questions on coping with shocks and on food security. Most of these questions were only added
at follow‐up. Table 3a reports that, at baseline, about 80 percent of households reported having
experienced a shock in the past 6 months, with the most common shock being a drought.
Table 3b displays the predicted poverty rates and we observe that poverty stood at over 40 percent.
Poverty rates in the study areas essentially remained unchanged between the baseline and follow‐up
survey with point estimates of 42.0 and 42.5 percent respectively (difference is not statistically
significant).15 In line with findings of previous studies and census data, poverty rates were highest in the
North‐East region (Karamoja) and lowest in the Mid‐North (around Gulu and Lira).
In accordance with these high poverty rates, we observe that (at follow‐up), only two‐thirds of Ugandans
in the rural North had ever attended school and in 45.9 percent of households, no household members
had completed at least primary education (table 3c). Only 22.7 percent of households had access to
electricity and a large increase can be observed between the baseline and follow‐up survey which is likely
associated with the spread of solar panels in rural Northern Uganda. Access to improved drinking water
was less of a challenge for rural dwellers in Northern Uganda with access rates of 80.5 percent. The vast
14 Almost all other transfers were hand or bus delivered by somebody in the households or a friend. 15 Poverty estimates were calculated using the most recent poverty line for rural Northern Uganda based on Uganda National Household Survey (UNHS) 2016‐17.
12
majority of rural dwellers were engaged in agricultural activities: 91.6 percent of households indicated
that they engaged in agriculture in the last 12 months and 51.0 percent owned livestock. Furthermore,
43.5 percent of households experienced a situation in which they did not have enough food in the past 12
months and 55.9 percent of households have a very low food security index. Despite high needs and
vulnerability, only 11.0 percent of households think they are able to come up with emergency cash16 if
needed.
Table 4 shows the poverty rates at follow‐up for households with different characteristics. Mobile money
users were on average less poor than non‐users: The poverty rate among mobile money users was 35
percent while that of non‐users was 45 percent. This relationship is hardly surprising given that usage of
mobile money requires access to a mobile phone, which is an expensive good, though we do find slightly
smaller differences in mobile phone ownership. Similarly, households which sent or received remittances
were less likely to be poor.
Households who have had the opportunity to save (with or without mobile money), who did not have to
take out loans, and have not had to incur a shock in the past 6 months, had lower poverty rates. Similarly,
households with low food security and no opportunity to access cash in case of an emergency, were about
7 percentage points more likely to be poor.
5.2 Impact on Usage and Financial Behavior
5.2.a. Impact on Usage and Financial Measures from Survey Data
We start by examining the effects of the Airtel Money agent rollout on mobile money usage. Table 5
reports the effects of the rollout on a dummy variable that is equal to one if the respondent said they
used mobile money in the last 3 months. We find no effect on this variable in the full sample (panel A of
table 2), but the heterogenous treatment effect regressions in panel B of table 3 show that the agent
rollout significantly increased mobile money usage by 4.2 percentage points in clusters that are far from
a bank branch, compared to a control group usage rate of 12.6 percent. There is no statistically significant
effect of the agent rollout on mobile money usage in clusters that are near a bank branch. These areas
already had slightly better access to mobile money agents before the roll‐out, with an average distance
to the closest agent of 6.6km, compared to 9.8km for areas far from a bank branch. The finding here
suggests that there is demand for mobile money in remote areas if services are made available.
Columns 2 and 3 of table 3 examine the effect of the agent rollout on savings. Households may increase
their overall savings or shift savings held in other places into mobile money accounts since mobile money
provides a relatively safe and easily accessible way of storing money. In fact, over two‐thirds of
respondents who save, report keeping the money on their person or at home, which may be less safe
compared to mobile money. However, we do not find an effect of the agent rollout on savings.
In the follow‐up survey, about 97 percent of mobile money users report that they use mobile money to
receive or send money, followed by 41 percent who use it to buy airtime. Only 4 percent say they use
mobile money in their business and 3 percent say they use it to pay bills. In tables 6a and 6b, we thus
16 Households were asked the following question: “Imagine that you have an emergency and you need to pay 126,000 shillings. How possible is it that you could come up with 126,000 shillings within the next month?”
13
investigate the effect of the agent rollout on receiving or sending money. We find no significant effect of
the rollout on the probability of receiving or sending money.
Tables 6a and 6b also report effects on the characteristic of the transfers for the person with whom
households transacted with most frequently. The effects on receiving or sending money via mobile money
(as opposed to other channels, such as hand carrying the money) are positive for clusters far from a bank
branch, but they are not statistically significant. While we find no effect on the frequency of sending
money, column 4 of table 3b shows a positive effect of the rollout on the probability of sending an amount
over 100 shillings for clusters that are far from a bank branch. In addition, column 7 of table 6a shows that
the rollout increased the probability of not paying any transportation cost for receiving money by 13
percentage points, compared to 69 percent in the control group, for clusters far away from a bank branch.
This finding likely reflects the fact that the agent rollout brought agents closer to the treatment clusters
and thus allowed respondents to walk or bike to the agent instead of having to pay for a motorcycle taxi
or mini‐bus taxi. The median transportation cost paid by respondents who had to pay this cost was 4,000
shillings at follow‐up in clusters far from a bank branch. This amount corresponds to about 10 percent of
daily household expenditures (median of 32,859 shillings, with a mean of 41,995 shillings).
Overall, although the agent rollout lowered the transportation costs of receiving money, table 6a and 6b
do not show that it increased the likelihood or frequency of receiving or sending money. It may be the
case that these effects occur for transacting with more people outside the household, which is not
captured in the survey data since it only asked for transactions with one person, i.e. the person
respondents received money from or sent money to most frequently. In section 5.2.b, we thus also
examine the effect of the agent money rollout on transactions recorded in Airtel’s administrative data.
Table 7 reports effects of the agent rollout on problems encountered when receiving or sending money.
In the full sample, respondents in treatment clusters are 5.7 percentage points less likely to say that the
agent was far, relative to 17.8 percent reporting the agent was far in control clusters. This effect is driven
by clusters far from a bank branch, where the agent rollout reduced the percentage of respondents saying
the agent was far by 14.5 percentage points (relative to 21.6 percent). For clusters far from a bank, the
agent rollout also reduced the percentage of respondents reporting high prices or fees by 10.5 percentage
points (relative to 28.1 percent in the control group). The results in table 7 are consistent with those in
table 6a that show that the agent rollout reduced transportation costs for receiving money in clusters far
from a bank branch.
In summary, we find that the agent rollout increased the probability of using mobile money in clusters
that are far away from a bank branch. We do not see an effect of the rollout on savings or the probability
of receiving or sending money as reported in the survey data. However, there is some evidence that
respondents in treatment clusters that are far away from a bank branch sent larger amounts of money
than those in control clusters and they pay lower transportation costs for receiving money. Finally,
compared to the control group, respondents who received or sent money in treatment clusters far away
from a bank branch are less likely to report that agents were far and that prices were high. Overall, the
agent rollout seems to have facilitated receiving and sending money in clusters far from a bank branch.
5.2.b. Impact on Transactions in Administrative Data
In table 8, we study the effect of the agent rollout on transactions in Airtel’s mobile money data. The
analysis here is at the cluster level. As described in section 4.2, we matched the transactions data to our
14
study clusters using the phone numbers reported by follow‐up survey respondents. The outcome
variables are dummies that are equal to one if the cluster had any phone number conducting the
respective transactions listed in columns 1 through 8 of table 8.
We do not find a statistically significant effect of the agent rollout on cash‐in, cash‐out, bill pay or top‐up
transactions. However, the agent roll‐out led to a significant increase in the probability of sending a P2P
transfer in the full sample (panel A of table 8) and in the probability of receiving a P2P transfer in clusters
far from a bank branch (panel B of table 8). Here, treatment clusters are 6.3 percentage points more likely
to have received a P2P transfer compared to control clusters. That is, in 9.3 percent of treatment clusters
received a P2P transfer in 2017 compared to only 3 percent of control clusters.
We did not find a statistically significant effect on the probability of sending or receiving P2P transfers in
the survey data (section 5.1), perhaps because the survey only asked about the individual the household
transacted with the most and thus does not capture all transfers. In fact, Jack, Ray, and Suri (2013) show
that remittances sent through M‐PESA come from more different senders, implying that it is important to
also consider results that include such transfers, as we do here with the transactions data.
The data here also have a caveat, namely that we see a relatively low number of clusters transacting,
which is in part due to the fact that we have a limited number of Airtel phone numbers in our survey data.
We are thus not capturing all Airtel mobile money transactions. Still, the results in table 8 provide
suggestive evidence that the agent rollout increased the probability of sending and receiving remittance
transfers, which is in line with the previous literature on mobile money agents (Blumenstock, Eagle and
Fafchamps 2016, Jack and Suri 2014, and Munyegera and Matsumoto 2016).
5.3 Impact on Labor Outcomes
Since our study was conducted in rural areas, we first examine whether the agent rollout had an impact
on households’ investment in livestock or agriculture. Greater ease of receiving and sending money may
have provided them with more liquidity for making these investments.17 Table 9 shows no effect of the
agent rollout on the probability that the household owns livestock or has engaged in agriculture.18 We
also do not see an effect on the number of different crops grown, the amount spent on seeds, or the
probability of using fertilizer.
We now ask whether the agent rollout affected labor decisions. Column 2 in table 10 shows that the agent
rollout increased the percentage of respondents who report doing non‐farm work by 2.8 percentage
points in clusters far from a bank branch, with no effect on clusters near a bank branch. The results in
column 3 further illustrate that the effect on non‐farm work is driven by self‐employment. That is,
respondents in treatment clusters far from a bank branch are 3 percentage points more likely to report
that they are self‐employed doing non‐farm work than respondents in control clusters. This effect is
17 To the extent that mobile money provides insurance against risk through increased remittance transactions, it can also provide incentives for increasing agricultural expenditures (see Karlan et al. 2014). This channel may not apply in our context though since we find no effect of the agent rollout on the probability or frequency of remittance transactions. 18 Due to a faulty skip‐pattern, questions about agriculture were only asked if respondents reported they owned livestock.
15
statistically significant and is almost equivalent to a doubling in the percentage of respondents doing such
work, i.e. going from 3.4 percent in control clusters to 6.4 percent in treatment clusters.
It thus appears that the agent rollout has allowed households to move from farm jobs into non‐farm self‐
employment. A possible channel is that increased liquidity from remittances or cost‐savings associated
with remittance transactions allowed them to make this switch. The costs savings accrue to 13 percent of
households who receive remittances, which in turn represents about 10 percent of households, giving
only 1.3 percent of households. However, for these households the cost savings are substantial (10
percent of daily consumption, as calculated in section 5.2). About 56 percent of remittance receivers
received a remittance only once in the last 6 months, 24 percent twice and 20 percent more than three
times.
Another possible channel is that increased self‐employment is caused directly by becoming Airtel agents.
Although the agent rollout selected pre‐existing businesses as agents, adding the mobile money service
may have increased business survival. To check the feasibility of this channel, we merged the phone
numbers that respondents provided with phone numbers of agents in Airtel’s transactions data. We do
not find any matches, suggesting that our survey respondents are not Airtel agents and the rollout thus
did not have direct employment effects on them.19
5.4 Impact on Food Security
Based on the previous literature that has shown that mobile money can contribute to coping with shocks
and consumption smoothing (Blumenstock, Eagle and Fafchamps 2016; Jack and Suri 2014), we examine
the effect of the agent rollout on responses to shocks and food security. As shown in the last row of
column 1 in table 11, 67.8 percent of respondents in control clusters reported that they experienced a
shock in the last 6 months. This number is not significantly different in treatment clusters. The most
commonly experienced shock was a drought (52.5 percent), followed by a flood (25.3 percent) and theft
or fire on the respondent’s property (17.3 percent).20
For respondents who reported experiencing a shock, table 11 examines the likelihood of using different
coping strategies. The outcome variables in columns 2 through 8 are dummy variables that are equal to 1
if the respondents reported using the coping strategy listed in the column title and equal to zero
otherwise. The most commonly used coping strategies were seeking help from local family, friends or
government and seeking help from individuals that live far away, followed by relying on savings or selling
assets (see control group means in table 11). The agent rollout did not have a significant impact on using
these three coping strategies. However, it decreased the likelihood of households changing their diet to
deal with shocks and increased the likelihood of households taking on work to deal with shocks. These
effects are driven by clusters far from a bank branch. In these clusters, respondents were 10.6 percentage
points less likely to adjust their diet to deal with a shock, compared to 23.5 percent in control clusters.
19 It could, however, have indirect effects if other businesses located near agents see an increase in sales due to people using agents ‐ either because people may use mobile money to pay for goods and services or because people may purchase goods and services during the trip they make to the agent. We do not have the necessary data on business activity to examine this channel more explicitly. 20 Other shocks elicited in the survey were, in order of reported frequency: having to contribute to a funeral; illness or accident; increase in fuel price; interest rate increase; death; theft, fire or loss of vehicle; and job loss.
16
Similarly, respondents were 5.1 percentage points more likely to take work, compared to only 8.9 percent
in control clusters.
To examine consumption smoothing further, the follow‐up survey included a module on food security.
Table 12 shows strong effects of the agent rollout on food security measures, which were again driven by
clusters far from a bank branch. In these clusters, the agent rollout reduced the fraction of respondents
that reported that they had to reduce the number of meals during the past 7 days by 11.8 percentage
points from a control group mean of 48.6 percent. Similarly, the fraction of respondents who had a very
low food security index dropped by 15.7 percentage points, compared to a control group mean of 62.9
percent. We find no effect of the agent rollout on food security in clusters near a bank branch.
In summary, we find that the agent rollout helped households in clusters far from a bank branch deal with
shocks by taking work instead of having to adjust their diet. In the same vein, the agent rollout increased
households’ food security and reduced the likelihood of having to skip meals. These results are consistent
with the previous finding that the agent rollout increased non‐farm self‐employment.
5.5 Impact on Poverty
Finally, we examine the effect of the agent rollout on households’ poverty status in table 13. Here, the
outcome variable is a dummy variable equal to 1 if the household was poor as estimated based on the
SWIFT methodology and equal to zero otherwise. The results show that the agent rollout did not have a
significant impact on being poor. This lack of an effect does not change with the distance of the cluster to
the nearest bank branch.
In summary, even though in section 5.1 (table 4), we observe that households who used mobile money
were better off than those who do not engage in mobile financial services, we do not find a causal effect
of the agent rollout on poverty levels, at least in the short‐run. A households’ poverty status is likely to
take more time to change than labor market outcomes or food security, so it is possible that the agent
rollout would reduce poverty levels in the longer‐run.
6. Conclusion
This paper studies the impact of rolling out mobile money agents in rural areas of Uganda. Compared to
areas included in previous impact evaluations of mobile money agents, the areas here tend to be poorer,
have lower access to financial services through bank branches and have almost no pre‐existing mobile
money agents.
In this setting, we find that the agent rollout significantly increased the percentage of households using
mobile money in areas further than 25.2km from a bank branch, leading to cost savings for those sending
or receiving remittances. Administrative data on mobile money transactions also suggest that the agent
rollout increased the probability of sending and receiving P2P transfers.
We do not find an effect of the agent rollout on savings or agricultural outcomes. However, in areas far
from a bank branch, the agent rollout statistically significantly increased the fraction of survey
respondents who work in non‐farm self‐employment. This increase appears to be driven by the cost
savings from remittance transactions and the increase in P2P transfer receipts, which households may use
to invest in self‐employment.
17
We also observe that the agent rollout changed the way in which households respond to negative shocks,
such as a drought or flood. The rollout increased the probability of taking work to cope with the shock and
it reduced the probability of changing the household’s diet. In line with this result, we find a strong positive
effect of the agent rollout on food security. However, we do not find an effect of the agent rollout on our
measure of poverty.
In terms of numbers, the estimates suggest that the rollout of 121 agents provided self‐employment to
257 households and improved food security for 1,345 households in areas far from a bank branch.21
References
Allen, Franklin, Elena Carletti, Robert Cull, Jun “Q. J.” Qian, Lemma Senbet, and Patricio Valenzuela. 2013. “Improving Access to Banking: Evidence from Kenya.” World Bank Policy Research Working Paper 6593. Blumenstock, Joshua E., Nathan Eagle and Marcel Fafchamps. 2016. “Airtime Transfers and Mobile Communications: Evidence in the Aftermath of Natural Disasters.” Journal of Development Economics 120: 157‐181. Bank of Uganda. 2017. Financial Stability Report 2017. Available at: https://www.bou.or.ug/bou/publications_research/Financial_Stability_Report.html Bruhn, Miriam, and Inessa Love. 2014. “The Real Impact of Improved Access to Finance: Evidence from Mexico.” The Journal of Finance, 69(3): 1347‐1376. Brune, Lasse, Xavier Giné, Jessica Goldberg, and Dean Yang. 2016. “Facilitating Savings for Agriculture: Field Experimental Evidence from Malawi.” Economic Development and Cultural Change 64(2): 187‐220. Burgess, Robin, and Rohini Pande. 2005. “Do Rural Banks Matter? Evidence from the Indian Social Banking Experiment.” American Economic Review 95(3): 780‐795. Buri, Sinja, Robert Cull, Xavier Giné, Sven Harten, and Soren Heitmann. 2018. “Banking with Agents: Experimental Evidence from Senegal.” World Bank Policy Research Working Paper 8417. CIPESA. 2018. National Information Technology survey: 2017/18 Report. Available at: https://www.nita.go.ug/sites/default/files/publications/National%20IT%20Survey%20April%2010th.pdf Dupas, Pascaline, and Jonathan Robinson. 2013. “Savings Constraints and Microenterprise Development: Evidence from a Field Experiment in Kenya.” American Economic Journal: Applied Economics, 5(1): 163–192. Dupas, Pascaline, Sarah Green, Anthony Keats, and Jonathan Robinson. 2016. “Challenges in Banking the
Rural Poor: Evidence from Kenya’s Western Province.” In S. Edwards, S. Johnson, D. Weil. eds. National
21 Calculations based on 8,576 households living in treatment clusters that are far from a bank branch, multiplied by 3 percent for the effect on self‐employment and by 15.7 percent for the effect on food security.
18
Bureau of Economic Research Conference Report. African Successes, Volume 3: Modernization and
Development Chicago: University of Chicago Press.
Dupas, Pascaline, Dean Karlan, Jonathan Robinson and Diego Ubfal. 2018. “Banking the Unbanked?
Evidence from three countries.” American Economic Journal: Applied Economics 10 (2): 257‐297.
Jack, William, Adam Ray, and Tavneet Suri. 2013. “Money Management by Households and Firms in
Kenya.” American Economic Review: Papers and Proceedings 103(3): 1‐8.
Jack, William and Tavneet Suri. 2014. “Risk Sharing and Transactions costs: Evidence from Kenya’s
Mobile Money Revolution.” American Economic Review 104(1): 183‐223.
Karlan, Dean, Roberto Darko Osei, Isaac Osei‐Akoto, and Christopher Udry. 2014. “Agricultural Decisions
after Relaxing Credit and Risk Constraints.” The Quarterly Journal of Economics 129(2): 597‐652.
Munyegera, Ggombe Kasim and Tomoya Matsumoto. 2016. “Mobile Money, Remittances, and
Households Welfare: Panel Evidence from Rural Uganda.” World Development 79: 127‐137.
Prina, Silvia. 2015. “Banking the Poor via Savings Accounts: Evidence from a Field Experiment.” Journal
of Development Economics, 115: 16‐31.
Sekabira, Haruna, and Matin Qaim. 2017. “Mobile Money, Agricultural Marketing, and Off‐Farm Income
in Uganda.” Agricultural Economics 48: 597‐611.
Suri, Taveneet, and William Jack. 2016. “The Long‐Run Poverty and Gender Impacts of Mobile Money.”
Science 354(6317): 1288‐1292.
Uganda Bureau of Statistics (UBOS). 2018. Uganda National Household Survey 2016/2017. Kampala,
Uganda. UBOS.
Yang, Dean, and HwaJung Choi. 2007. “Are Remittances Insurance? Evidence from Rainfall Shocks in the
Philippines.” World Bank Economic Review 21(2): 219‐248.
Young, Nathaniel. 2015. “Formal Banking and Economic Growth: Evidence from a Regression
Discontinuity Analysis in India.” Mimeograph, Boston University.
19
Figure 1: Number of Active Agents and Usage
Panel A: Number of Active Agents by month Panel B: Number of Transactions by month
Source: Own calculations based on Airtel Money transaction data.
Figure 2: Areas Included in the Study
Note: The map shows 334 clusters of enumeration areas that were included in the study. These areas were
assigned priority ratings 3 or 4 by Airtel. Areas with priority ratings 1 and 2 received few agents and are dropped in
the analysis. Randomization into treatment and control groups was stratified by priority level.
0
50
100
150
200
250
300
Acitve Agents (at least 1 transaction)
‐
5,000
10,000
15,000
20,000
25,000
30,000
Number of transactions for active agents
20
Figure 3: Example of Treatment and Control Areas with Agents
Note: The mat shows treatment and control clusters in the Gulu district. Study agents are Airtel agents that were
rollout out as part of this study. Pre‐study agents are mostly MTN agents that existed in 2015.
21
Table 1: Mobile Phone Penetration in Uganda
Areas in Uganda
Urban Rural Rural North
(1) (2) (3)
Owns a mobile phone (%) 78.5 65.7 27.6
Has access to a mobile phone (%) 39.9
Service provider for phone owners is (%) MTN 73.7 74.7
Airtel 61.4 7.9
Africell 7.3 9.9
Note: Data in columns 1 and 2 from CIPESA (2018). National Information Technology survey: 2017/18 Report. Data in column 3 from 2018 follow‐up survey.
22
Table 2a: Airtel Data on Mobile Money Transactions
Average number of transactions across
clusters
Dummy for clusters with non‐zero transactions
2016 2017 2016 2017
Send P2P transfer 0.350 0.677 0.069 0.066
Receive P2P transfer 0.749 1.338 0.075 0.090
Cash‐in 2.132 3.955 0.120 0.123
Cash‐out 2.302 3.323 0.126 0.120
Bill pay 0.410 1.395 0.027 0.057
Airtime top‐up 9.416 13.746 0.099 0.123
Data top‐up 0.120 0.449 0.045 0.057
All 15.479 24.883 0.138 0.147
Note: Data for 334 clusters. Airtel transactions data was matched to study clusters based on phone numbers of survey respondents. The table does thus not include most Airtel mobile money transactions, which cannot be mapped to our study locations.
Table 2b: Airtel Data on Mobile Money Transactions (2016)
Treatment group mean
Control group mean
P‐value of difference in means
Dummy for clusters with non‐zero transactions (1) (2) (3)
Send P2P transfer 0.077 0.060 0.522
Receive P2P transfer 0.083 0.066 0.536
Cash‐in 0.125 0.114 0.743
Cash‐out 0.119 0.133 0.734
Bill pay 0.024 0.030 0.712
Airtime top‐up 0.107 0.090 0.605
Data top‐up 0.036 0.054 0.425
Any transaction 0.137 0.139 0.980
Note: Data for 334 clusters. Airtel transactions data was matched to study clusters based on phone numbers of survey respondents. The table does thus not include most Airtel mobile money transactions, which cannot be mapped to our study locations. Column 3 shows the p‐value of the difference in treatment and control group means, clustered at the cluster level and conditional on strata dummies.
23
Panel A: Baseline survey data (1) (2) (3) (4) (5) (6)
Household head is male 3638 0.71 0.75 0.081
Household head's age 3636 42.51 14.79 41.65 14.38 0.192
No household member completed primary 3710 0.46 0.44 0.45
Household size 3720 5.38 3.04 5.54 3.06 0.312
Has access to a phone 3720 0.43 0.46 0.423
Used mobile money in last 3 months 3720 0.14 0.14 0.605
Saved money in last 6 months 3720 0.78 0.79 0.503
Saved money in mobile account 3720 0.06 0.06 0.629
Received money in the last 6 months 3720 0.15 0.15 0.976
Received money via mobile money 557 0.42 0.46 0.454
Received money more than once 557 0.58 0.53 0.491
Avg. amount received over 100 shilling 557 0.33 0.37 0.526
Paid no fees for receiving money 557 0.32 0.28 0.593
Fee for receiving money (if not zero) 391 3342.12 6391.88 3596.39 7864.14 0.692
Paid no transp costs for receiving money 557 0.55 0.53 0.74
Transp costs for receiving money (if not zero) 256 9746.46 27868.97 5808.53 12752.35 0.135
Sent money in the last 6 months 3720 0.11 0.12 0.379
Sent money via mobile money 429 0.54 0.49 0.455
Sent money more than once 429 0.55 0.57 0.701
Avg. amount sent over 100 shilling 429 0.34 0.4 0.341
Paid no fees for sending money 429 0.22 0.25 0.423
Fee for sending money (if not zero) 326 9438.24 41563.1 3773.43 9088.3 0.149
Paid no transp costs for sending money 429 0.48 0.53 0.558
Transp costs for sending money (if not zero) 211 5442.26 5310.31 7094.31 11121.63 0.129
High prices/fees 805 0.32 0.3 0.458
Agent was far 805 0.2 0.17 0.516
Technical problem 805 0.2 0.22 0.862
Charged wrong amount 805 0.1 0.14 0.113
Has job that earns income 3720 0.6 0.66 0.104
Does non‐farm work 3720 0.11 0.13 0.147
Self‐employed (non‐farm) 3720 0.05 0.07 0.144
Experienced shock in last 6m 3720 0.79 0.8 0.582
Panel B: Follow‐up survey data
Household head is male 4469 0.71 0.75 0.02
Household head's age 4461 44.56 15.12 44.64 15.17 0.838
No household member completed primary 4541 0.46 0.45 0.634
Household size 4384 5.57 3.19 5.76 3.17 0.261
Note: Column 6 shows the p‐value of the difference in treatment and control group means, clustered at the cluster
level and conditional on strata dummies.
Table 3a: Summary Statistics
Mean MeanStandard
deviation
Control group Treatment group
Standard
deviation
Number of
households
P‐value of
difference in
means
24
Table 3b: Imputed Poverty Rates at Baseline and Follow‐up
Panel A: Rural North
Poverty rate Confidence interval lower bound
Confidence interval upper bound
Baseline 42.0 39.6 44.5
Follow‐up 42.5 39.9 45.2
Panel B: By region
Poverty rate Confidence interval lower bound
Confidence interval upper bound
Mid‐North Baseline 35.9 32.7 39.0
North East Baseline 64.4 58.5 70.3
West Nile Baseline 42.9 39.4 46.4
Mid‐North Follow‐up 39.2 35.8 42.7
North East Follow‐up 64.7 58.9 70.5
West Nile Follow‐up 39.5 36.1 43.0
Note: Estimates based on survey‐to‐survey imputation technique (SWIFT) using poverty line of UNHS 2016‐17 for rural Northern Uganda
Table 3c: Household Characteristics Over Time
Baseline Follow‐up
Household size 5.4 5.4
HoH ever attended school 68.1 63.6
HoH completed primary education 45.0 45.9
HH uses electricity for lighting 7.8 22.7
Access to improved water 77.3 80.5
HH engaged in agriculture 91.6
HH engaged in non‐farm work 11.7 8.6
HH owns livestock 51.0
HH experienced food shortage 43.5
HH has very low food security 55.9
HH has access to emergency cash 14.2 11.0
Note: HH=household; HoH= head of household
25
Poverty rateConfidence interval
lower bound
Confidence interval
upper bound
Phone ownership: no 44.5 41.7 47.4
Phone ownership: yes 37.3 33.2 41.5
Ever used mobile money: no 45.0 42.1 47.9
Ever used mobile money: yes 35.1 31.2 39.0
Received mobile money: no 42.6 31.5 53.6
Received mobile money: yes 34.1 29.9 38.4
Sent mobile money: no 38.1 29.3 46.9
Sent mobile money: yes 34.4 30.1 38.6
Saved in past 6 months: no 48.3 44.3 52.3
Saved in past 6 months: yes 40.8 37.9 43.7
Saved in past 6 months using mobile money: no 43.6 40.9 46.3
Saved in past 6 months using mobile money: yes 29.3 22.2 36.5
Taken loans in past 6 months: no 42.9 40.2 45.7
Taken loans in past 6 months: yes 40.4 35.2 45.5
Experienced shock in past 6 months: no 40.2 36.8 43.5
Experienced shock in past 6 months: yes 43.7 40.7 46.8
Very low food security: no 38.9 36.0 41.7
Very low food security: yes 45.4 42.1 48.8
Could get emergency cash: no 43.4 40.6 46.1
Could get emergency cash: yes 35.9 30.4 41.4
Table 4: Poverty Rates by MM Characteristics at Follow‐Up
Note: Estimates based on survey‐to‐survey imputation technique (SWIFT) using poverty line of UNHS 2016‐17
for rural Northern Uganda
26
Panel A:
ITT for all clusters (1) (2) (3)
Agent ‐0.002 ‐0.017 ‐0.012
(0.016) (0.027) (0.010)
R‐squared 0.042 0.026 0.024
N 4469 4469 4469
Number of clusters 334 334 334
Control group mean 0.180 0.771 0.079
Panel B:
ITT by distance from bank branch
Agent*Far 0.042* 0.002 ‐0.015
(0.022) (0.043) (0.014)
Agent*Near ‐0.035 ‐0.031 ‐0.009
(0.021) (0.035) (0.015)
R‐squared 0.045 0.027 0.024
N 4469 4469 4469
Number of clusters 334 334 334
Control group mean ‐ Far 0.126 0.716 0.061
Control group mean ‐ Near 0.218 0.809 0.091
Table 5: Effects of Agent Rollout on Usage and Financial Behavior
Note: Panel A shows results from OLS regressions of the dependent variables on a
dummy variable that is equal to 1 if the cluster was randomly selected for the agent
rollout and equal to zero otherwise. Panel B shows results from similar regressions,
where the agent rollout dummy is interacted with two dummy variables indicating the
distance to the nearest Airtel partner bank branch, where far is more than 25.2km away
from a branch and near is less than or equal to 25.2km away from a branch. All
regressions include randomization strata dummies and a dummy for whether the
household head is male. When data on the outcome was collected at baseline, we also
control for the baseline value of the outcome variable. For observations where the
baseline value of the outcome is missing, we replace this value with zero and include a
dummy variable indicating that the value is missing. Control group means are means of
the dependent variables. Standard errors clustered at the cluster level in parenthesis.
Statistical significance levels: * p<0.10, ** p<0.05, *** p<0.01.
Used mobile
money in last 3
Saved money
in last 6
Saved money in
mobile account
Dependent variable:
27
Panel A: ITT for all clusters (1) (2) (3) (4) (5) (6) (7) (8)
Agent 0.010 0.016 0.023 0.040 ‐0.042 0.127 0.083** ‐0.054
(0.015) (0.045) (0.045) (0.044) (0.041) (0.121) (0.038) (0.391)
R‐squared 0.029 0.039 0.009 0.025 0.075 0.040 0.038 0.138
N 4469 606 606 606 420 345 604 122
Number of clusters 334 215 215 215 189 170 215 83
Control group mean 0.131 0.736 0.468 0.281 0.192 7.541 0.755 8.037
Panel B: ITT by distance from bank branch
Agent*Far 0.013 0.051 ‐0.060 0.005 ‐0.105 0.274 0.130** 0.141
(0.017) (0.076) (0.081) (0.068) (0.065) (0.244) (0.065) (0.439)
Agent*Near 0.008 ‐0.001 0.062 0.056 ‐0.010 0.054 0.062 ‐0.175
(0.022) (0.057) (0.054) (0.056) (0.052) (0.133) (0.046) (0.541)
R‐squared 0.029 0.040 0.012 0.026 0.078 0.043 0.039 0.140
N 4469 606 606 606 420 345 604 122
Number of clusters 334 215 215 215 189 170 215 83
Control group mean ‐ Far 0.094 0.693 0.477 0.261 0.238 7.569 0.690 8.123
Control group mean ‐ Near 0.157 0.754 0.464 0.290 0.170 7.529 0.783 7.986
Table 6a: Effects of Agent Rollout on Receiving Money
Dependent variable:
Note: Panel A shows results from OLS regressions of the dependent variables on a dummy variable that is equal to 1 if the cluster was randomly selected for
the agent rollout and equal to zero otherwise. Panel B shows results from similar regressions, where the agent rollout dummy is interacted with two dummy
variables indicating the distance to the nearest Airtel partner bank branch, where far is more than 25.2km away from a branch and near is less than or equal to
25.2km away from a branch. All regressions include randomization strata dummies and a dummy for whether the household head is male. When data on the
outcome was collected at baseline, we also control for the baseline value of the outcome variable. For observations where the baseline value of the outcome
is missing, we replace this value with zero and include a dummy variable indicating that the value is missing. Control group means are means of the dependent
variables. Standard errors clustered at the cluster level in parenthesis. Statistical significance levels: * p<0.10, ** p<0.05, *** p<0.01.
Received
money in the
last 6 months
Received
money via
mobile money
Received
money more
than once
Avg. amount
received over
100 shilling
Paid no fees
for receiving
money
Log fee for
receiving
money
Paid no transp
costs for
receiving money
Log transp costs
for receiving
money
28
Panel A: ITT for all clusters (1) (2) (3) (4) (5) (6) (7) (8)
Agent ‐0.013 0.021 0.025 0.073 ‐0.042 0.135 0.021 0.107
(0.014) (0.054) (0.048) (0.048) (0.039) (0.155) (0.055) (0.199)
R‐squared 0.027 0.059 0.022 0.039 0.035 0.039 0.033 0.133
N 4469 493 493 493 493 414 497 129
Number of clusters 334 191 191 191 191 181 193 78
Control group mean 0.115 0.688 0.477 0.342 0.185 7.717 0.725 8.147
Panel B: ITT by distance from bank branch
Agent*Far ‐0.001 0.121 ‐0.027 0.134* ‐0.049 0.036 0.090 0.065
(0.024) (0.085) (0.076) (0.074) (0.053) (0.197) (0.078) (0.416)
Agent*Near ‐0.023 ‐0.042 0.059 0.034 ‐0.038 0.196 ‐0.023 0.128
(0.017) (0.069) (0.060) (0.062) (0.055) (0.217) (0.072) (0.220)
R‐squared 0.027 0.066 0.024 0.041 0.035 0.039 0.037 0.134
N 4469 493 493 493 493 414 497 129
Number of clusters 334 191 191 191 191 181 193 78
Control group mean ‐ Far 0.096 0.678 0.444 0.300 0.178 7.546 0.725 8.203
Control group mean ‐ Near 0.129 0.694 0.494 0.365 0.188 7.809 0.725 8.117
Note: Panel A shows results from OLS regressions of the dependent variables on a dummy variable that is equal to 1 if the cluster was randomly selected for
the agent rollout and equal to zero otherwise. Panel B shows results from similar regressions, where the agent rollout dummy is interacted with two dummy
variables indicating the distance to the nearest Airtel partner bank branch, where far is more than 25.2km away from a branch and near is less than or equal to
25.2km away from a branch. All regressions include randomization strata dummies and a dummy for whether the household head is male. When data on the
outcome was collected at baseline, we also control for the baseline value of the outcome variable. For observations where the baseline value of the outcome
is missing, we replace this value with zero and include a dummy variable indicating that the value is missing. Control group means are means of the dependent
variables. Standard errors clustered at the cluster level in parenthesis. Statistical significance levels: * p<0.10, ** p<0.05, *** p<0.01.
Table 6b: Effects of Agent Rollout on Sending Money
Dependent variable:
Sent money in
the last 6
months
Sent money
via mobile
money
Sent money
more than
once
Avg. amount
sent over 100
shilling
Paid no fees
for sending
money
Log fee for
sending money
Paid no transp
costs for sending
money
Log transp costs
for sending
money
29
Panel A: ITT for all clusters (1) (2) (3) (4)
Agent ‐0.044 ‐0.057* ‐0.009 0.006
(0.032) (0.030) (0.024) (0.024)
R‐squared 0.040 0.017 0.024 0.031
N 867 867 867 867
Number of clusters 245 245 245 245
Control group mean 0.259 0.178 0.120 0.090
Panel B: ITT by distance from bank branch
Agent*Far ‐0.105** ‐0.145*** ‐0.044 0.004
(0.052) (0.047) (0.045) (0.041)
Agent*Near ‐0.012 ‐0.010 0.010 0.007
(0.041) (0.039) (0.027) (0.029)
R‐squared 0.042 0.025 0.025 0.031
N 867 867 867 867
Number of clusters 245 245 245 245
Control group mean ‐ Far 0.281 0.216 0.180 0.101
Control group mean ‐ Near 0.249 0.160 0.092 0.085
Note: Panel A shows results from OLS regressions of the dependent variables on a dummy
variable that is equal to 1 if the cluster was randomly selected for the agent rollout and
equal to zero otherwise. Panel B shows results from similar regressions, where the agent
rollout dummy is interacted with two dummy variables indicating the distance to the
nearest Airtel partner bank branch, where far is more than 25.2km away from a branch and
near is less than or equal to 25.2km away from a branch. All regressions include
randomization strata dummies and a dummy for whether the household head is male.
When data on the outcome was collected at baseline, we also control for the baseline value
of the outcome variable. For observations where the baseline value of the outcome is
missing, we replace this value with zero and include a dummy variable indicating that the
value is missing. Control group means are means of the dependent variables. Standard
errors clustered at the cluster level in parenthesis. Statistical significance levels: * p<0.10,
** p<0.05, *** p<0.01.
Dependent variable:
Problems encountered when receiving or sending money
Table 7: Effects of Agent Rollout on Problems with Receiving or Sending Money
High
prices/feesAgent was far
Technical
problem
Charged wrong
amount
30
Panel A: ITT for all clusters (1) (2) (3) (4) (5) (6) (7) (8)
Agent ‐0.019 0.048* 0.036 0.005 ‐0.012 ‐0.019 ‐0.007 ‐0.006
(0.038) (0.027) (0.031) (0.036) (0.035) (0.025) (0.036) (0.025)
R‐squared 0.054 0.020 0.028 0.047 0.038 0.035 0.049 0.024
N 334 334 334 334 334 334 334 334
Control group mean 0.157 0.042 0.072 0.120 0.127 0.066 0.127 0.060
Panel B: ITT by distance from bank branch
Agent*Far ‐0.016 0.037 0.063* 0.011 0.012 0.011 0.023 0.012
(0.047) (0.033) (0.037) (0.044) (0.044) (0.028) (0.043) (0.028)
Agent*Near ‐0.023 0.057 0.011 0.000 ‐0.034 ‐0.045 ‐0.034 ‐0.023
(0.059) (0.042) (0.049) (0.055) (0.054) (0.041) (0.055) (0.041)
R‐squared 0.054 0.020 0.030 0.047 0.040 0.039 0.051 0.025
N 334 334 334 334 334 334 334 334
Control group mean ‐ Far 0.103 0.021 0.030 0.077 0.077 0.026 0.064 0.026
Control group mean ‐ Near 0.205 0.057 0.114 0.159 0.170 0.102 0.182 0.091
Note: Analysis at the cluster level. Airtel transactions data for 2017 that was matched to the 334 study clusters based on phone numbers of survey
respondents. Panel A shows results from OLS regressions of the dependent variables on a dummy variable that is equal to 1 if the cluster was
randomly selected for the agent rollout and equal to zero otherwise. Panel B shows results from similar regressions, where the agent rollout dummy
is interacted with two dummy variables indicating the distance to the nearest Airtel partner bank branch, where far is more than 25.2km away from a
branch and near is less than or equal to 25.2km away from a branch. All regressions include randomization strata dummies. Control group means are
means of the dependent variables. Standard errors clustered at the cluster level in parenthesis. Statistical significance levels: * p<0.10, ** p<0.05, ***
p<0.01.
Table 8: Effects of Agent Rollout on Airtel Mobile Money Transactions (2017 Administrative Data)
Dependent variable (dummy=1 if any phone number in the cluster made the following transactions):
Any type of
transaction
Send P2P
transfer
Receive P2P
transferCash‐in Cash‐out Bill pay
Airtime top‐
upData top‐up
31
Panel A: ITT for all clusters (1) (2) (3) (4) (5)
Agent 0.012 ‐0.020 0.018 ‐3922.246 ‐0.001
(0.034) (0.016) (0.190) (10302.628) (0.009)
R‐squared 0.014 0.010 0.026 0.010 0.009
N 4469 2275 2084 2084 2084
Number of clusters 334 307 293 293 293
Control group mean 0.501 0.925 3.870 67971.260 0.029
Panel B: ITT by distance from bank branch
Agent*Far 0.039 ‐0.027 ‐0.417 ‐6523.055 0.003
(0.054) (0.030) (0.313) (14458.055) (0.013)
Agent*Near ‐0.008 ‐0.015 0.303 ‐2217.289 ‐0.004
(0.044) (0.017) (0.228) (14238.775) (0.013)
R‐squared 0.015 0.010 0.033 0.010 0.009
N 4469 2275 2084 2084 2084
Number of clusters 334 307 293 293 293
Control group mean ‐ Far 0.457 0.909 4.049 67120.319 0.026
Control group mean ‐ Near 0.532 0.934 3.763 68476.627 0.031
Note: Panel A shows results from OLS regressions of the dependent variables on a dummy variable that is
equal to 1 if the cluster was randomly selected for the agent rollout and equal to zero otherwise. Panel B
shows results from similar regressions, where the agent rollout dummy is interacted with two dummy
variables indicating the distance to the nearest Airtel partner bank branch, where far is more than 25.2km
away from a branch and near is less than or equal to 25.2km away from a branch. All regressions include
randomization strata dummies and a dummy for whether the household head is male. When data on the
outcome was collected at baseline, we also control for the baseline value of the outcome variable. For
observations where the baseline value of the outcome is missing, we replace this value with zero and
include a dummy variable indicating that the value is missing. Control group means are means of the
dependent variables. Standard errors clustered at the cluster level in parenthesis. Statistical significance
levels: * p<0.10, ** p<0.05, *** p<0.01.
Table 9: Effects of Agent Rollout on Agricultural Outcomes
Dependent variable:
Owns
livestock
Has engaged in
agriculture in
last 12 months
Number of
different crops
Amount spent
on seeds Uses fertilizer
32
Panel A: ITT for all clusters (1) (2) (3)
Agent ‐0.023 0.004 0.01
‐0.032 ‐0.012 ‐0.009
R‐squared 0.027 0.031 0.013
N 4469 4463 4463
Number of clusters 334 334 334
Control group mean 0.337 0.083 0.042
Panel B: ITT by distance from bank branch
Agent*Far ‐0.025 0.028* 0.030**
(0.046) (0.016) (0.012)
Agent*Near ‐0.022 ‐0.013 ‐0.006
(0.044) (0.017) (0.012)
R‐squared 0.027 0.033 0.015
N 4469 4463 4463
Number of clusters 334 334 334
Control group mean ‐ Far 0.305 0.068 0.034
Control group mean ‐ Near 0.360 0.094 0.048
Note: Panel A shows results from OLS regressions of the dependent variables on a
dummy variable that is equal to 1 if the cluster was randomly selected for the
agent rollout and equal to zero otherwise. Panel B shows results from similar
regressions, where the agent rollout dummy is interacted with two dummy
variables indicating the distance to the nearest Airtel partner bank branch, where
far is more than 25.2km away from a branch and near is less than or equal to 25.2km
away from a branch. All regressions include randomization strata dummies and a
dummy for whether the household head is male. When data on the outcome was
collected at baseline, we also control for the baseline value of the outcome
variable. For observations where the baseline value of the outcome is missing, we
replace this value with zero and include a dummy variable indicating that the value
is missing. Control group means are means of the dependent variables. Standard
errors clustered at the cluster level in parenthesis. Statistical significance levels: *
p<0.10, ** p<0.05, *** p<0.01.
Table 10: Effects of Agent Rollout on Occupation
Dependent variable:
Has job that
earns income
Does non‐farm
work
Self‐employed
(non‐farm)
33
Panel A: ITT for all clusters (1) (2) (3) (4) (5) (6) (7) (8)
Agent ‐0.024 ‐0.024 ‐0.007 0.019 ‐0.064** 0.015 0.029* 0.024
(0.032) (0.035) (0.039) (0.032) (0.028) (0.023) (0.017) (0.026)
R‐squared 0.051 0.013 0.037 0.043 0.026 0.026 0.006 0.029
N 4469 2970 2970 2970 2970 2970 2970 2970
Number of clusters 334 313 313 313 313 313 313 313
Control group mean 0.678 0.681 0.509 0.324 0.230 0.129 0.097 0.212
Panel B: ITT by distance from bank branch
Agent*Far ‐0.043 ‐0.079 ‐0.046 0.009 ‐0.106*** 0.012 0.051* ‐0.000
(0.052) (0.055) (0.057) (0.048) (0.038) (0.029) (0.027) (0.038)
Agent*Near ‐0.009 0.014 0.020 0.026 ‐0.035 0.017 0.014 0.041
(0.040) (0.045) (0.054) (0.043) (0.040) (0.033) (0.022) (0.035)
R‐squared 0.051 0.015 0.038 0.043 0.028 0.026 0.006 0.029
N 4469 2970 2970 2970 2970 2970 2970 2970
Number of clusters 334 313 313 313 313 313 313 313
Control group mean ‐ Far 0.662 0.705 0.515 0.306 0.235 0.085 0.089 0.190
Control group mean ‐ Near 0.690 0.664 0.505 0.336 0.226 0.160 0.102 0.227
Note: Panel A shows results from OLS regressions of the dependent variables on a dummy variable that is equal to 1 if the cluster was randomly selected for
the agent rollout and equal to zero otherwise. Panel B shows results from similar regressions, where the agent rollout dummy is interacted with two dummy
variables indicating the distance to the nearest Airtel partner bank branch, where far is more than 25.2km away from a branch and near is less than or equal to
25.2km away from a branch. All regressions include randomization strata dummies and a dummy for whether the household head is male. When data on the
outcome was collected at baseline, we also control for the baseline value of the outcome variable. For observations where the baseline value of the outcome
is missing, we replace this value with zero and include a dummy variable indicating that the value is missing. Control group means are means of the dependent
variables. Standard errors clusterd at the cluster level in parenthesis. Statistical significance levels: * p<0.10, ** p<0.05, *** p<0.01.
Coped with shock through
Table 11: Effects of Agent Rollout on Coping with Shocks
Dependent variable:
Experienced
shock in last
6m
Local help Help from
faraway
Relying on
savings or
selling assets
Changing diet Obtaining
credit Taking work Other
34
Panel A: ITT for all clusters (1) (2)
Agent ‐0.067** ‐0.075**
(0.033) (0.033)
R‐squared 0.037 0.047
N 4469 4469
Number of clusters 334 334
Control group mean 0.488 0.598
Panel B: ITT by distance from bank branch
Agent*Far ‐0.118** ‐0.157***
(0.046) (0.048)
Agent*Near ‐0.028 ‐0.014
(0.045) (0.044)
R‐squared 0.039 0.052
N 4469 4469
Number of clusters 334 334
Control group mean ‐ Far 0.486 0.629
Control group mean ‐ Near 0.489 0.576
Table 12: Effects of Agent Rollout on Food Security
Dependent variable:
Had to reduce the
number of meals in
the last 7 days
Food security
index is very low
Note: Panel A shows results from OLS regressions of the dependent variables on a
dummy variable that is equal to 1 if the cluster was randomly selected for the agent
rollout and equal to zero otherwise. Panel B shows results from similar regressions,
where the agent rollout dummy is interacted with two dummy variables indicating
the distance to the nearest Airtel partner bank branch, where far is more than 25.2km
away from a branch and near is less than or equal to 25.2km away from a branch. All
regressions include randomization strata dummies and a dummy for whether the
household head is male. When data on the outcome was collected at baseline, we
also control for the baseline value of the outcome variable. For observations where
the baseline value of the outcome is missing, we replace this value with zero and
include a dummy variable indicating that the value is missing. Control group means
are means of the dependent variables. Standard errors clustered at the cluster level
in parenthesis. Statistical significance levels: * p<0.10, ** p<0.05, *** p<0.01.
35
Panel A: ITT for all clusters Dependent variable: Poverty status
Agent ‐0.003
(0.021)
R‐squared .
N 4469
Number of clusters 334
Control group mean 0.428
Panel B: ITT by distance from bank branch
Agent*Far 0.008
(0.030)
Agent*Near 0
(0.026)
R‐squared .
N 4469
Number of clusters 334
Control group mean ‐ Far 0.421
Control group mean ‐ Near 0.432
Note: Panel A shows results from OLS regressions of the dependent variables on a
dummy variable that is equal to 1 if the cluster was randomly selected for the agent
rollout and equal to zero otherwise. Panel B shows results from similar regressions,
where the agent rollout dummy is interacted with two dummy variables indicating the
distance to the nearest Airtel partner bank branch, where far is more than 25.2km away
from a branch and near is less than or equal to 25.2km away from a branch. All regressions
include randomization strata dummies and a dummy for whether the household head is
male. When data on the outcome was collected at baseline, we also control for the
baseline value of the outcome variable. For observations where the baseline value of the
outcome is missing, we replace this value with zero and include a dummy variable
indicating that the value is missing. Control group means are means of the dependent
variables. Standard errors clustered at the cluster level in parenthesis. Statistical
significance levels: * p<0.10, ** p<0.05, *** p<0.01.
Table 13: Effects of Agent Rollout on Poverty